Where humans must remain in control of AI agents

Agentic AI changes the risk conversation because the system may do more than generate an answer. It may retrieve records, update data, trigger workflows or communicate externally.

Human control should match consequence

The National AI Centre recommends meaningful human oversight and clear intervention points that allow people to pause, override, roll back or shut down AI systems. That does not mean every low-risk internal action needs a senior executive click. It means approval is deliberately placed where the potential harm, cost or loss of trust is material.

Three useful action classes

• Prepare: the agent drafts, summarises, classifies or recommends, but a person decides what happens next.
• Execute within limits: the agent performs a reversible, low-risk action under defined rules and logs the result.
• Human approval required: the agent cannot complete an external, financial, legal, employment, safety, production or sensitive-data action without an accountable person.

Design the exception path first

A good agent knows when not to continue. Low confidence, missing information, unusual data, conflicting instructions, policy limits or a failed integration should create an exception with an owner. Silent failure is more dangerous than an obvious refusal.

Oversight is a role, not a button

The reviewer needs sufficient context to make a real decision: the request, source material, proposed action, uncertainty, relevant policy and expected impact. A generic “approve” button without this evidence creates theatre rather than control.

Keep a manual path

Critical processes must continue if an AI service is unavailable or withdrawn. Document the alternative workflow, preserve data portability and periodically test the ability to pause automation. Resilience is part of responsible adoption, particularly as more work becomes dependent on external model providers.